Loading skill
Install any skill in seconds. Free to start, no credit card required.
Get Started Free →Strategy: AI-safety systematic probing — enumerate all threat surfaces, generate attack vectors per surface, execute probes, and aggregate findings across the full attack space.
.claude/skills/yogsoth-ai-systematic-probing/SKILL.md| Test case | Without → With | Effect | Δ tokens | Δ turns |
|---|---|---|---|---|
| case-12 | ✗→✓ | ▲ Improved | 2% | 0% |
| case-04 | ✗→✓ | ▲ Improved | 10% | 0% |
| case-05 | ✗→✓ | ▲ Improved | 53% | 0% |
| case-06 | ✗→✓ | ▲ Improved | -49% | 0% |
| case-09 | ✗→✓ | ▲ Improved | 27% | 0% |
Anthropic-style systematic probing: exhaustive coverage of all threat surfaces with structured attack generation and execution.
| Parameter | S | M | L | |---|---|---|---| | Attack vectors | 5 | 12 | 20 | | Probing rounds | 3 | 6 | 10 | | Personas | 2 | 4 | 6 | | Assumption checks | 5 | 10 | 20 |
threat-surface-mapping → [enumerate surfaces]
→ [for each surface]:
attack-vector-generation (generate vectors)
→ [for each vector]:
probe-execution (execute attack)
→ (if partial success: generate follow-up vectors)
→ finding-aggregation → attack-resilience-scoring<!-- BEGIN available-tables (generated) -->
Optional, no fixed order; the final leaf is always a sop.
| Tactic | When to use | | --- | --- | | assumption-cascade | Tactic: Surface assumptions, sort by dependency, attack root assumptions first, then trace cascade failures through the dependency graph. | | structured-attack-campaign | Tactic: Full attack lifecycle — threat surface enumeration, attack vector generation, systematic probing, and finding aggregation across all surfaces. |
Optional, no fixed order; the final leaf is always a sop.
| SOP | When to use | | --- | --- | | attack-resilience-scoring | Compute overall resilience score (0.0-1.0) based on attack results, coverage, and vulnerability severity distribution. | | attack-vector-generation | Generate specific attack strategies for a given threat surface, producing concrete probes that can be executed. | | finding-aggregation | Aggregate, deduplicate, and classify findings from multiple probes into a coherent vulnerability report. | | probe-execution | Execute a single attack probe against an artifact, record the result with evidence and severity classification. | | threat-surface-mapping | Enumerate all attackable surfaces of an artifact — logical, empirical, methodological, social, and practical dimensions. |
<!-- END available-tables (generated) -->
| Case | Status | Duration (ms) | Turns | Tokens | Tool calls | ||||||||
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| Without | With | Δ | Without | With | Δ | Without | With | Δ | Without | With | Δ | ||
case-12 | fail→pass | 12,415 | 15,053 | +21% | 1 | 1 | 0% | 2,048 | 2,099 | +2% | 0 | 0 | — |
case-01 | fail→fail | 15,166 | 14,947 | -1% | 1 | 1 | 0% | 1,118 | 1,781 | +59% | 0 | 0 | — |
case-02 | fail→fail | 26,809 | 32,103 | +20% | 1 | 1 | 0% | 2,890 | 5,381 | +86% | 0 | 0 | — |
case-03 | fail→fail | 21,141 | 28,754 | +36% | 1 | 1 | 0% | 2,164 | 4,580 | +112% | 0 | 0 | — |
case-04 | fail→pass | 15,325 | 18,143 | +18% | 1 | 1 | 0% | 2,416 | 2,664 | +10% | 0 | 0 | — |
case-05 | fail→pass | 12,516 | 25,471 | +104% | 1 | 1 | 0% | 1,456 | 2,233 | +53% | 0 | 0 | — |
case-06 | fail→pass | 25,066 | 16,154 | -36% | 1 | 1 | 0% | 3,084 | 1,576 | -49% | 0 | 0 | — |
case-13 | pass→pass | 22,464 | 21,647 | -4% | 1 | 1 | 0% | 2,228 | 3,235 | +45% | 0 | 0 | — |
case-07 | pass→pass | 17,470 | 14,202 | -19% | 1 | 1 | 0% | 2,510 | 2,543 | +1% | 0 | 0 | — |
case-08 | fail→fail | 21,425 | 17,270 | -19% | 1 | 1 | 0% | 2,694 | 2,754 | +2% | 0 | 0 | — |
case-09 | fail→pass | 18,521 | 17,936 | -3% | 1 | 1 | 0% | 2,852 | 3,617 | +27% | 0 | 0 | — |
case-10 | fail→pass | 19,644 | 17,541 | -11% | 1 | 1 | 0% | 1,644 | 3,292 | +100% | 0 | 0 | — |
case-11 | pass→pass | 15,573 | 19,129 | +23% | 1 | 1 | 0% | 1,637 | 2,820 | +72% | 0 | 0 | — |
case-14 | fail→pass | 21,776 | 15,393 | -29% | 1 | 1 | 0% | 2,436 | 2,953 | +21% | 0 | 0 | — |
case-15 | fail→pass | 12,866 | 16,138 | +25% | 1 | 1 | 0% | 1,258 | 2,206 | +75% | 0 | 0 | — |
case-16 | fail→fail | 19,555 | 13,027 | -33% | 1 | 1 | 0% | 2,274 | 1,153 | -49% | 0 | 0 | — |
case-17 | pass→pass | 18,203 | 21,157 | +16% | 1 | 1 | 0% | 2,847 | 3,810 | +34% | 0 | 0 | — |
case-18 | fail→pass | 21,386 | 5,479 | -74% | 1 | 1 | 0% | 2,594 | 974 | -62% | 0 | 0 | — |
case-19 | pass→pass | 7,664 | 8,134 | +6% | 1 | 1 | 0% | 1,356 | 1,963 | +45% | 0 | 0 | — |
case-20 | pass→pass | 16,329 | 17,009 | +4% | 1 | 1 | 0% | 2,646 | 3,214 | +21% | 0 | 0 | — |
case-21 | pass→pass | 13,605 | 11,919 | -12% | 1 | 1 | 0% | 1,617 | 2,774 | +72% | 0 | 0 | — |
case-22 | fail→pass | 21,495 | 17,665 | -18% | 1 | 1 | 0% | 2,613 | 1,020 | -61% | 0 | 0 | — |
DecimalAI ran this skill against gemini-3.6-flash twice over the same eval suite — once with the skill loaded and once without — and compared the two runs case by case. 22 cases were attempted. The headline lift of +45 percentage points is the difference between those two pass rates over the 22 comparable cases.
Without the skill loaded, the model failed this case. With it loaded, the same prompt on the same model passed. This is one improved case from the latest verified run; every case, including any that regressed, is in the table above.
Other measured skills in the registry, with their headline benchmark lift.